id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_13800 | When a more specific answer is present in the document, generic references have been treated as incorrect. | sometimes there is not a more specific reference; for example an article written before a drug has been released may never name the drug. | contrasting |
train_13801 | We find that precision numbers against Freebase are low, below 10%. | these numbers are not reliable mainly because many correct instances found by our models are missing in Freebase. | contrasting |
train_13802 | Supervised learning can discover more general patterns (Kambhatla, 2004;Culotta and Sorensen, 2004). | this approach requires labeled data, and most work only carry out experiments on small data set. | contrasting |
train_13803 | Researchers also use topic models to perform dimension reduction on features when they cluster relations (Hachey, 2009). | they do not explicitly model entity types. | contrasting |
train_13804 | The learning algorithm is actually fast enough to do this automatically after each labeling action. | we found such a dynamically changing interface to be frustrating for users (e.g., words they wanted to label would move or disappear). | contrasting |
train_13805 | This is essentially the common feature-selection method for identifying the most salient features in text classification (Sebastiani, 2002). | we use both L and probabilistically-labeled instances from U to compute IG(f k ), to better reflect what the model believes it has learned. | contrasting |
train_13806 | These pathological cases represent potential pitfalls that could be alleviated with additional user studies and training. | we note that the active dual interface is not particularly worse in these cases, it is simply not significantly better, as in the other 13 trials. | contrasting |
train_13807 | We present the complete approach in three parts by describing the factored representation of the lexicon (Section 5), techniques for proposing potential new lexemes and templates (Section 6), and finally a complete learning algorithm (Section 7). | the next section first reviews the required background on semantic parsing with CCG. | contrasting |
train_13808 | For this reason, UBL never proposes the lexical item, cheapest NP\(S|NP)/(S|NP) : λ f λ g.argmin(λ x. f (x) ∧ g(x), λ y.cost(y)), which is used to parse the sentence in Figure 2. | fUBL uses a lexeme learned from the same word in different contexts, along with a template learnt from similar words in a similar context, to learn to per- figure 2: An example learned parse. | contrasting |
train_13809 | The results in Table 3 are similar to Banko and Etzioni's findings that a set of eight POS patterns cover a large fraction of binary verbal relation phrases. | their analysis was based on a set of sentences known to contain either a company acquisition or birthplace relationship, while our results are on a random sample of Web sentences. | contrasting |
train_13810 | For AL, the choice as to which labels are used (as a result of voting, bypassing or other) also has an influence on the selection. | we had to keep the sequence of the selected sentences fixed in the simulations reported above. | contrasting |
train_13811 | It can be argued that the use of short listings simplifies the problem of attribute extraction, since short listings can be easily annotated and one can apply supervised learning approach to extract product attributes. | as the size of the data grows, obtaining labeled training set on the scale of millions of listings becomes very expensive. | contrasting |
train_13812 | The web context is subsequently used to extract additional automobile brands, which result in a total of 5701 brands. | the reported results in (Nadeau el al 2006) have low precision, in some case less than 50%. | contrasting |
train_13813 | Because an influenza patient usually requires anti-influenza drugs, this approach is reasonable. | in most countries, antiinfluenza drugs are not available at the drug store (only hospitals provide such drugs). | contrasting |
train_13814 | From Table 4, we observe that NB, CRM and TAM, as generative models, tend to suggest general tags such as "novel", "literature", "classic" and "France", and fail in suggesting specific tags such as "Alexandre Dumas" and "Count of Monte Cristo". | wTM succeeds in suggesting both general and specific tags related to the book. | contrasting |
train_13815 | Finally, we find that user history can be a good indicator of rumors. | we believe that this feature could be more helpful with a complete user set and a more comprehensive history of their activities. | contrasting |
train_13816 | These feature schemas are not the only possible ones-they were empirically selected for the specific purpose of augmenting the Charniak parser. | much subsequent work has tended to use these same features, albeit sometimes with extensions for specific purposes (e.g. | contrasting |
train_13817 | In Chambers's work, discriminative classifiers -maximum entropy (MaxEnt) classifiers were used by incorporating linguistic features and temporal constraints for training, which outperforms the previous temporal language models on a subset of Gigaword Corpus (Graff et al., 2003). | the conventional methods have some limitations because they predict creation time of documents mainly based on feature-based models without understanding content of documents, which may lead to wrong predictions in some cases. | contrasting |
train_13818 | In this case, it is difficult to recognize the relative temporal relations. | timeliness can be leveraged to determine the relations as a heuristic method. | contrasting |
train_13819 | For example, triple r d,ex =< d, ex, −1 > means that the time of extraction ex is one year before the time of document d. Extractions from different documents have no connections. | there are a great number of extractions referring to the same event. | contrasting |
train_13820 | When only 1,000 documents are initially labeled with timestamps, the confidence boosting model can propagate their timestamps to more than 400,000 documents with an accuracy of 0.494, approximately 12.8% relative improvement over the BFS counterpart, which proves effectiveness of the confidence boosting model. | as shown in table 5, hardly can the propagation process propagate timestamps to all documents. | contrasting |
train_13821 | In addition, they used KL-divergence instead of normalized log likelihood ratio to measure differences between a document and a time period's language model. | these methods are based on temporal language models so they also suffer from the problem of the method of de Jong et al. | contrasting |
train_13822 | Given only syntactic features, we may be drawn to conclude that they share similar temporal relationships. | in the first line of text, the events temporally OVERLAP, while in the second line they do not. | contrasting |
train_13823 | If the relations are randomly distributed, we should expect their distribution to follow that of the temporal classes as shown in Table 1. | we see that many of the relations do not follow this distribution. | contrasting |
train_13824 | This part of the model is parametric, operating over a fixed number of tags T , and is identical to the formulation of tag transitions in the Bayesian HMM (Goldwater and Griffiths, 2007). | we replace the BHMM's emission distribution with the morphologically-informed distributions described below. | contrasting |
train_13825 | We find no improvement on tagging performance in English when adding morphology, compared to the MORTAGNOSEG baseline in which words are not segmented. | we do see a small but significant improvement over the BHMM for both of these models, due to the replacement of the multinomial emission distribution in the BHMM with the PYP. | contrasting |
train_13826 | Conversely, in Spanish we saw less difference on the morphology task between models with categories inferred solely from morphemic patterns and models that also used local syntactic context for categorisation. | in Spanish we saw an improvement in the tagging task when morphology information was included. | contrasting |
train_13827 | (2012) present a model that is similar to ours, using a noisy channel model implemented with a finite-state transducer to learn about phonetic variability while clustering distinct tokens into lexical items. | (like the earlier lexical-phonetic learning model of Feldman et al. | contrasting |
train_13828 | The unigram model is a reasonable segmenter, though not quite as good as the bigram model, with boundary F of 76.4 and token F of 60.9 (compared to 79.8 and 64.4 using the bigram model). | it is not good at normalizing variation; its mtk score is comparable to the baseline at 44.8% 5 . | contrasting |
train_13829 | Because there is little available training data, and because social media language changes rapidly (Eisenstein, 2013b), fully supervised training is generally not considered appropriate for this task. | due to the extremely high-dimensional output space -arbitrary sequences of words across the vocabulary -it is a very challenging problem for unsupervised learning. | contrasting |
train_13830 | We are left with a difference in expected feature counts, as is typical in log-linear models. | unlike the supervised case, here both terms are expectations: the outer expectation is over all target sequences (given the observed source sequence), and the nested expectation is over all source sequences, given the target sequence. | contrasting |
train_13831 | It may appear that all we have gained by applying sequential Monte Carlo is to convert a computational problem into a statistical one: a naive sampling approach will have little hope of finding the small high-probability region of the highdimensional label space. | sequential importance sampling allows us to address this issue through the proposal distribution, from which we sample the candidate words t n . | contrasting |
train_13832 | Microblogs such as Twitter, Sina Weibo (a popular Chinese microblog service) and Facebook have received increasing attention in diverse research communities (Han and Baldwin, 2011;Hawn, 2009, inter alia). | to traditional text domains that use carefully controlled, standardized language, microblog content is often informal, with less adherence to conventions regarding punctuation, spelling, and style, and with a higher proportion of dialect or pronouciation-derived orthography. | contrasting |
train_13833 | If we look at the normalization DanielVeuleman to Daniel Veuleman, we can see that it is only applicable when the exact word DanielVeuleman occurs. | we wish to learn that it is uncommon for the letters l and v to occur in the same word sequentially, so that be can add missing spaces in words that contain the lv character sequence, such as normalizing phenomenalvoter to phenomenal voter. | contrasting |
train_13834 | However, we wish to learn that it is uncommon for the letters l and v to occur in the same word sequentially, so that be can add missing spaces in words that contain the lv character sequence, such as normalizing phenomenalvoter to phenomenal voter. | there are also cases where this is not true, for instance, in the word velvet, we do not wish to separate the letters l and v. Thus, we shall describe the process we use to decide when to apply these transformations. | contrasting |
train_13835 | This includes orthographic errors, abbreviations and slang. | this is generally not enough to detect lexical variants, as many words share similar contexts, such as already, recently and normally. | contrasting |
train_13836 | This can be captured by allowing each speaker to have a different θ and ζ in each stage of the campaign. | we expect that a speaker draws from his ideological lexicon similarly across different epochs-there is a single ψ shared between different epochs. | contrasting |
train_13837 | In MIX, there are no temporal effects between cue terms, although the structure of our ideology tree encourages the speaker to draw from coarse-grained ideologies over fine-grained ideologies. | the strong Markovian dependency between states in NORES would encourage the model to stay local within the ideology tree. | contrasting |
train_13838 | The parameters of the token based ITGs can be estimated using expectation maximization through an efficient dynamic programming algorithm in conjunction with beam pruning (Saers and Wu, 2011). | to token based ITGs, each rule in a segmental ITG grammar can contain more than one token in both input and output languages. | contrasting |
train_13839 | It is not immediately obvious how one might quantify such a notion of "idea-based" influence. | the mechanism used in the scientific community for giving credit to prior work is citation. | contrasting |
train_13840 | We will show that such an assumption is quite reasonable under our model, since the prototype projections successfully tease out the proper semantics from these aggregate representations. | it is natural to wonder whether one can do better if one incorporates the compositional model into the training of the word representations in the first place. | contrasting |
train_13841 | This result shows that our learning model successfully captures good representation within cocompositionality of additive model. | to other previous compositional models, our model does not require estimating a large number of parameters for computation of compositional vectors and word representation itself is more suitable for it. | contrasting |
train_13842 | The appeal of this kind of simple approach is its intuitive geometric interpretation and its robustness to various datasets. | it may not be sufficiently expressive to represent the various factors involved in compositional semantics, such as syntax and context. | contrasting |
train_13843 | Many of these models need to optimize n × n parameters, which may be large. | our model only needs two hyper-parameters: the number of prototype words m and dimensional reduction k in SVD (Table 6). | contrasting |
train_13844 | For example, in the phrase creative new idea, the idea is both new and creative, so we would expect a similar impact of modification by both adjectives. | we predict that in rigid order cases, one adjective, the one closer to the noun, will dominate the meaning of the phrase, distorting the meaning of the noun by a significant amount. | contrasting |
train_13845 | To understand the message-passing algorithm in this study, it is helpful to think of our model as a simplified Hidden Semi-Markov Model (HSMM), in which the letters represent the states and the speech features are the observations. | unlike in a regular HSMM, where the state sequence is hidden, in our case, the state sequence is fixed to be the given letter sequence. | contrasting |
train_13846 | By comparing our model to the grapheme baseline, we can see the advantage of modeling the pronunciations of a letter using a mixture model, especially for a language like English which has many pronunciation irregularities. | even for languages with straightforward pronunciation rules, the concept of modeling letter pronunciations using mixture models still applies. | contrasting |
train_13847 | Previous work on machine comprehension (e.g., semantic modeling) has made great strides, but primarily focuses either on limited-domain datasets, or on solving a more restricted goal (e.g., open-domain relation extraction). | mCTest requires machines to answer multiple-choice reading comprehension questions about fictional stories, directly tackling the high-level goal of open-domain machine comprehension. | contrasting |
train_13848 | Yet these techniques are necessarily evaluated individually, rather than by how much they advance us towards the end goal. | the goal of semantic parsing is the machine comprehension of text (MCT), yet its evaluation requires adherence to a specific knowledge representation, and it is currently unclear what the best representation is, for open-domain text. | contrasting |
train_13849 | The intent was to produce a set of stories and questions that varied in difficulty so that research work can progress grade-by-grade if needed. | we found little difference between grades in the corpus.. After gathering the stories, we manually curated the MC160 corpus by reading each story set and correcting errors. | contrasting |
train_13850 | ShefLM achieves a high compression ratio when it stores counts of ngrams in the training corpus. | when this method stores probabilities of ngrams, the advantage of using compression is limited because floating-point numbers are difficult to compress. | contrasting |
train_13851 | In prior work, the endmarker symbol has been used to indicate whether an ngram is in the trie. | there is no need to distinguish whether the node of the tree is included in the language model because all nodes of a backwards suffix tree which represents ngrams surely exist in the model. | contrasting |
train_13852 | In this paper, we focus on penalized maximum likelihood estimation in log-linear models. | to language models based on unstructured norms such as 2 (quadratic penalties) or 1 (absolute discounting), we use tree-structured norms (Zhao et al., 2009;Jenatton et al., 2011). | contrasting |
train_13853 | The generalization properties of this algorithm is as good as the generalization obtained with the T 2 penalty, if not better. | it has the constant value non-branching path property, which Table 1: Properties of the algorithms proposed in this paper. | contrasting |
train_13854 | Phrase-based models have shown a very strong performance when translating between languages that have similar word orders. | they are not able to adequately capture the complex relationships that exist between the word orders of languages of different families such as English and Chinese. | contrasting |
train_13855 | A source span boundary feature in Figure 2(b) that is defined on the source parse tree is also a local feature. | a target span boundary feature in Figure 2(c), which assesses the target parse structure, is a non-local feature. | contrasting |
train_13856 | Non-local features enable us to model the target parse structure in a derivation. | it is computationally expensive to calculate the expected values of non-local features over D (s) words and result in an extremely large number of states during dynamic programming. | contrasting |
train_13857 | The metrics provide an effective overall ranking, as the systems with high scores generally make fewer errors. | the metrics do not convey the significant variation in the types of errors systems make. | contrasting |
train_13858 | Some of these cases are correct, such as variation in person between mentions inside and outside of quotes. | many of these cases are errors. | contrasting |
train_13859 | 4 By fixing each of the other error types in isolation, we can get a sense of the gain if just that error type is addressed. | it also means some mentions are incorrectly placed in the same cluster, causing some negative scores. | contrasting |
train_13860 | No single source of errors stands out as the most substantial challenge today. | it is worth noting that while confidence measures can be used to reduce precision-related errors, no system has been able to effectively address the recall-related errors, such as Missed Entities. | contrasting |
train_13861 | Unfortunately, it is difficult to compare directly to the results of both systems, since they reported results on portions of ACE and CoNLL datasets using gold mentions. | our approach provides independent evidence for the benefit of NEL, and joint modeling in particular, since it outperforms the state-of-the-art Stanford sieve system (winner of the CoNLL 2011 shared task (Pradhan et al., 2011)) and other recent comparable approaches on benchmark datasets. | contrasting |
train_13862 | One limitation of our approach is that it only learns the properties that are present in CSN antecedents. | aSN antecedents have additional properties which are not always captured by CSN antecedents. | contrasting |
train_13863 | We propose a scalable semi-supervised feature engineering approach. | to previous works using pre-defined taskspecific features with fixed values, we dynamically extract representations of label distributions from both an in-domain corpus and an out-of-domain corpus. | contrasting |
train_13864 | Therefore, MI can show the tendency of two strings forming one word. | previous works mainly focused on the balanced case, i.e., the MI of strings with the same length. | contrasting |
train_13865 | Note that the SSFs are still calculated using all the unlabeled data. | each iteration in the algorithm uses unlabeled data with different sizes. | contrasting |
train_13866 | Our interpretation is that the correct nominative reading is pruned from the 0-order lattice. | the higher-order models can put less weight on 0-order features as they have access to more context to disambiguate the sequence. | contrasting |
train_13867 | The sentences in the corpus have a relatively simple syntax, though many also exhibit long distance dependencies. | these conversations are pragmatically complex. | contrasting |
train_13868 | For the first baseline 34 examples are like this; for the second and third baselines it's 32. | this problem occurs only once with the fourth baseline, and all the trades predicted by our method are complete, making random choice unnecessary. | contrasting |
train_13869 | Previous work has shown that for many common NLP tasks, 7 Turkers' average score can match expert annotations (Snow et al., 2008). | we use 15 Turkers because we had no gold-standard data and because we were not sure how difficult the task is. | contrasting |
train_13870 | The best performing method for the Action genre is CP+Web Search at 87.67%, while the best performing method for the Romance genre is PMI+Web search at 88.52%. | pMI+Web Search does not beat Cp+Web Search on average over both genres we tested, even though the Mechanical Turk HIT for Cp specifies that the order of the events matters: a more stringent criterion. | contrasting |
train_13871 | In particular, coreference patterns tend to be very similar in translations of a text, and this fact has been exploited with good results to project coreference annotations from one language into another by using word alignments (Postolache et al., 2006;Rahman and Ng, 2012). | what is true in general need not be true for all types of linguistic elements. | contrasting |
train_13872 | Our empirical results have shown that this approach outperforms a previous graph-based approach with an absolute gain of 9%. | while these graph-based approaches perform effectively when the agent has perfect knowledge or perception of the environment (e.g., 84%), they perform rather poorly when the agent has imperfect perception of the environment (e.g., 45%). | contrasting |
train_13873 | The approach is a direct extension of the incremental algorithm (Dale, 1995). | this work only provides a proof of concept example to illustrate the idea. | contrasting |
train_13874 | A directed hypergraph G (Gallo et al., 1993) is a tuple of the form: Similar to regular graphs, a hypergraph consists of a set of nodes X and a set of arcs A. | different from regular graphs, each arc in A is considered as a hyperarc in the sense that it can capture relations between any two subsets of nodes: a tail (t i ) and a head (h i ). | contrasting |
train_13875 | On the one hand, we did not have a large set of human descriptions of the impoverished scene to learn the stochastic cost. | it is not clear whether human strategies of describing the impoverished scene should be used to represent optimal strategies for the robot. | contrasting |
train_13876 | This seems to indicate that exploring extra effort in REG could help mediate mismatched perceptions in situated dialogue. | more understanding on how to engage in such extra effort will be required in the future. | contrasting |
train_13877 | 1.00 1.00 0.80 0.30 fwd sports cars 1.00 1.00 1.00 1.00 garden landscaping magazines 0.00 0.10 0.15 0.06 heliskiing resorts 1.00 1.00 1.00 1.00 hell in a cell wrestlers 1.00 1.00 1.00 0.92 holidays celebrated in sydney Table 7: Precision at various ranks in the ranked lists of instances extracted from queries, for various target class labels and as an average over the entire set of 50 target class labels lore" and "daft punk live albums", and especially for "garden landscaping magazines" which has the worst precision. | instances extracted for "companies with sustainable competitive advantage" or "criminals who have been executed" have high precision across all ranks. | contrasting |
train_13878 | (Bunescu and Pasca, 2006) propose to generalize beyond context-entity correlation s(d, e) with word-category correlation s(w, c). | this method works at word level, and does not scale well to large number of categories. | contrasting |
train_13879 | (Cucerzan, 2007) first suggests to optimize an objective function that is similar to the collective ap-proach. | the author adopts an approximation method because of the large search space (which is O(n m ) for a document with m mentions, each with n candidates). | contrasting |
train_13880 | For example, consider the sentence: "It also marks P&G's growing concern that its Japanese rivals, such as Kao Corp., may bring their superconcentrates to the U.S." According to our annotation style, there is a relation "rivals" between "P&G" and "Kao Corp." in this sentence. | the original annotations for PENN-100 consider only the relation between "Kap Corp." and the pronoun "it", leaving the task of resolving the coreference between "P&G" and "it" as a posterior step. | contrasting |
train_13881 | Recently, a lot of work (Sarkas et al., 2010;Li et al., 2009) proposed the task of structured annotation of queries which aims to detect the structure of the query and assign a specific label to it. | to our knowledge, the previous methods do not exploit an effective approach for improving web search ranking by incorporating structured annotation of queries. | contrasting |
train_13882 | For the sake of simplicity, this paper assumes that attributes in domain schema are available. | it is not difficult to pre-specify attributes in a specific domain. | contrasting |
train_13883 | Figure 3(b) shows that the precision of the structured annotation is lowest when δ = 0 . | the precision is still as high as 0.7375, and the highest recall is obtained in this case. | contrasting |
train_13884 | Machine translation already has a decades long history and an array of commercial systems were already deployed, including Google Translate 1 and Systran 2 . | due to the intrinsic difficulty of the task, a number of related problems remain open, including: the gap between text semantics and statistically derived translations, the scarcity of resources in a large majority of languages and the quality of automatically obtained resources and translations. | contrasting |
train_13885 | This finding advocates for the importance of good quality generic dictionaries in specialized lexicon translation approaches. | it is clear that such dictionaries are far from being sufficient in order to cover all possible domains. | contrasting |
train_13886 | Most of the features are extracted based only on the sentence to be compressed. | we introduce a few document level features. | contrasting |
train_13887 | Increasing the order of our dominance model provides an additional gain. | the gain is more pronounced in the weblog genre (up to around 1 BLEU point) than in the newswire genre. | contrasting |
train_13888 | Even when the number of possible permutations is large we can limit ourselves to the K most popular permutations. | our approach provides important advantages. | contrasting |
train_13889 | With a triangulation pivot approach, a source-target phrase table can be obtained by combining the source-pivot phrase table and the pivot-target phrase table. | one of the weaknesses is that some corresponding source and target phrase pairs cannot be generated, because they are connected to different pivot phrases (Cui et al., 2013). | contrasting |
train_13890 | IHMM for system combination, HMM in GIZA++ software for statistical machine translation (SMT) (Och and Ney, 2000;Koehn et al., 2003), are designed for one-to-many alignment, and running GIZA++ from two directions to gain better performance turns into a standard operation in SMT, therefore we are seeking a way to empower IHMM by introducing bi-directional information. | it appears to be intractable in an IHMM model to search the optimal solution by simply defining a new goal as a product of probabilities from two directions. | contrasting |
train_13891 | The unnormalized conditional alignment probability is . | the same alignment (f 1 , e 1 )(f 1 , e 2 )(f 2 , e 3 ), if we change the alignment direction, the backbone being observations, would be a bit different. | contrasting |
train_13892 | A straightforward method to introduce awareness of source syntax to translation rules is to apply the same well-formed dependency constraint and head POS annotation on the target side of stringto-dependency translation rules to the source side. | as discussed earlier, this would significantly reduce the number of rules that can be extracted, exacerbate data sparsity, and cause other problems, especially given that the target side is already constrained by the dependency requirement. | contrasting |
train_13893 | Previous studies have exploited revision history data in tasks such as preposition error correction (Cahill et al., 2013), spelling error correction (Zesch, 2012) or paraphrasing (Max and Wisniewski, 2010). | they all use different approaches to extract the information needed for their task. | contrasting |
train_13894 | In edit category classification, we also have two documents. | these documents are different versions of the same text. | contrasting |
train_13895 | It is the most straightforward approach when dealing with multi-labeled data. | it does not consider possible relationships or dependencies between categories. | contrasting |
train_13896 | Or in the case of identical alignment, New York City aligning to New York City is simply New↔New, York↔York, City↔City. | it is not as clear how to token-align New York (as a city) with New York City. | contrasting |
train_13897 | Deng and Byrne (2008) explored token-to-phrase alignment based on HMM models (Vogel et al., 1996) by explicitly modeling the token-to-phrase probability and phrase lengths. | the token-to-phrase alignment is only in one direction: each target state still only spans one source word, and thus alignment on the source side is limited to tokens. | contrasting |
train_13898 | This strategy has been successful and commonly used in coreference resolution (Ng and Cardie, 2002;Bengtson and Roth, 2008;Stoyanov et al., 2009). | most works have developed ad-hoc approaches to implement this idea. | contrasting |
train_13899 | Another approach proposed by Yu and Joachims (2009) formu-lates coreference with latent spanning trees. | their approach has no directionality between mentions, whereas our latent structure captures the natural left-to-right ordering of mentions. | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.